脑影像智能分析与脑疾病早期诊断ppt课件.pptx

《脑影像智能分析与脑疾病早期诊断ppt课件.pptx》由会员分享，可在线阅读，更多相关《脑影像智能分析与脑疾病早期诊断ppt课件.pptx（78页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、脑影像智能分析与脑疾病 早期诊断张道强南京航空航天大学Brain Projects美国“脑活动图谱”计划欧盟“人类脑计划”中国脑计划Brain Imaging(Neuroimaging)Neuroimaging includes the use of various techniques to either directly or indirectly image the structure or function of the brainTwo broad categoriesStructural neuroimaging deals with the structure of the br

2、ainFunctional neuroimaging is used to indirectly measure brain functionsNeuroimaging-based Classification(S.Lemm,et al.,Neuroimage,2011)Example:Brain Decoding(Nature Feature News,2013)(T.Mitchell et al.,Science,2008)Recovery MoviesOutlineSummary123Backgrounds on Alzheimers DiseaseBrain-imaging based

3、 AnalysisBrain-network based Analysis4History of ADAD was first described by German psychiatrist and neuropathologist Alois Alzheimer in 1906 and was named after himThe 51 y.o.woman(Auguste Deter)cared by Dr.Alzheimer until her death in 1906.He did an autopsy,examined her brain&described the typical

4、 abnormalities of what would be called later Alzheimers DiseaseWhat Is AD?It is the most common form of dementiaThere is no cure for the disease,which worsens as it progresses,and eventually leads to deathMost often,AD is diagnosed in people over 65 years of ageIn 2006,there were 26.6 million suffer

5、ers worldwide,and it is predicted to affect 1 in 85 people globally by 2050不同年代我国痴呆不同年代我国痴呆和和A D 患患者的人数者的人数【Lancet.2013】1.“三三高高”：患病率高、患病率高、致残率高、致残率高、负负担重担重2.“三低三低”：就就诊诊率低、率低、诊诊断率低、治断率低、治疗疗率低率低目前我目前我国国A D 的患病率的患病率【Alzheimers&Dementia.2013】我我国国A D 现现状状:Celebrities with ADNormal vs.AD BrainIn the norma

6、l brain there is a lot of healthy brain tissue in the language area.In the AD affected brain there is little in that areaThere are many differences between the two brains including the memory,sulcus,gyrus,ventricle,and language areas.In the AD brain,these are either shrunken or stretched out to unhe

7、althy measuresNormal vs.AD BrainForms abnormal clumps called amyloid plaques and tangled bundles of fibers called neurofibrillary tangles in the brainAD自画像自画像1967（早年）（早年）1996（患病第（患病第2年）年）1997（患病第（患病第3年）年）199819992000Normal or diseased?(S.Crutch,et al.,Lancet Neurology,2012)Normal or diseased?(S.Crut

8、ch,et al.,Lancet Neurology,2012)AD ProgressionAD atrophy progresses Starts in the medial temporal and limbic areas Hippocampus and entorhinal cortex Subsequently spreading to parietal association areas Finally to frontal and primary corticesAD BiomarkersBiomarkers for early diagnosis of AD Magnetic

9、resonance imaging(MRI)Positron emission tomography(PET)Cerebrospinal fluid(CSF)-A42,t-tau and p-tau MRIPETCSFBiomarkersOutlineSummary123Backgrounds on Alzheimers DiseaseBrain-imaging based AnalysisBrain-network based Analysis4Multimodal ClassificationMotivation Several modalities of biomarkers have

10、been proved to be sensitive to AD,or its prodromal stage,i.e.,mild cognitive impairment(MCI)Different biomarkers provide complementary information,which may be useful for diagnosis of AD or MCI when used togetherQuestion:How can we effectively combine both imaging data(MRI and PET)and non-imaging da

11、ta(CSF)for multi-modality based classification?FlowchartTemplateMRIdataPETdataCSFdata68,131,21,42,FeatureextractionFeatureextractionFeature selectionCalculatekernel matrixCalculatekernel matrixCalculatekernel matrixSVMoptionalKKeerrnneell ccoommbbiinnaattiioonn(D.Zhang,et al.Neuroimage,2011)Material

12、s202 subjects from ADNI,including 51 AD patients,99 MCI and 52 healthy controls 43 MCI converters who had converted to AD within 18 monthsand 56 MCI non-converters who had not converted Only baseline data of MRI,CSF and PET are usedAD(n=51;18F/33M)MCI(n=99;32F/67M)HC(n=52;8F/34M)MeanSDRangeMeanSDRan

13、geMeanSDRangeAge75.27.459-8875.37.055-8975.35.262-85Education14.73.64-2015.92.98-2015.83.28-20MMSE23.82.020-2627.11.724-30291.225-30CDR0.70.30.5-10.50.00.5-0.500.00-0ResultsAD vs.HCMCI vs.HCMethodsACC(%)SEN(%)SPE(%)ACC(%)SEN(%)SPE(%)MRI86.2(82.9-89.0)86(82.7-88.7)86.3(83.1-89.1)72.0(68.4-74.7)78.5(7

14、5.6-80.6)59.6(55.1-63.7)CSF82.1(80-84.9)81.9(80-84.7)82.3(80-85.1)71.4(68.2-73.3)78(75.6-79.4)58.8(54.3-61.7)PET86.5(82.9-90.5)86.3(82.7-90.3)86.6(83.1-90.6)71.6(67.4-74.7)78.2(75-80.6)59.3(52.9-63.7)Combined93.2(89.0-96.5)93(88.7-96.3)93.3(89.1-96.6)76.4(73.5-79.7)81.8(79.4-84.4)66.0(62.6-70.3)Base

15、line91.5(88.5-96.5)91.4(88.3-96.3)91.6(88.6-96.6)74.5(71.9-78.2)80.4(78.3-83.3)63.3(59.7-68.3)Comparison of performance of single-modaland multimodal classification methods(D.Zhang,et al.Neuroimage,2011)Multi-Modal Multi-Task LearningMotivation Besides classification,there also exist regression task

16、s which estimate continuous clinical scores to evaluate the stage of AD pathology and predict future progression Both regression and classification tasks are essentially related due to the same underlying pathologyQuestion:How can we jointly predict multiple regression and classification variables f

17、rom multi-modality data?AD/MCI/HCMMSEADAS-Cog(D.Zhang,D.Shen.Neuroimage,2012)FlowchartTemplateMRIdataPETdataCSFdataFeatureextractionFeatureextractionSVM(Regression/Classification)68,131,21,42,ClinicalscoresMMSE；ADAS-CogClass LabelsCalculate kernel matrixSVM(Regression/Classification)Multi-modelSVMKK

18、eerrnneell combinationCalculatekkeerrnneell mmaattrriixxCCaallccuullaattee kernel matrixMTFSMulti-task feature selectionMTFS(D.Zhang,D.Shen.Neuroimage,2012)MaterialsADNI Subjects 186subjects(45AD,91 MCI and 50 HCs),only baseline data,3 modalities(MRI,CSF and PET)deviati ExperimentsExperiment 1 Estim

19、ating clinical stages MMSE,ADAS-Cog,and class label(AD/MCI/HC)Experiment 2 Predicting 2-year MMSE and ADAS-Cog changes and MCI conversionResultsComparison of performances of different methods on Experiment 1Results(contd)Comparison of performances of different methods on Experiment 2Manifold Regular

20、ized Multitask Learning(B.Jie,D.Zhang,et al.Human Brain Mapping,2015)Multi-level Multitask Learning(M.Wang,et al.MICCAI 2017)Longitudinal Multitask Learning(D.Zhang,et al.PLOS ONE,2012;B.Jie,et al.IEEE TBME 2017)Multimodal Transfer Learning(B.Cheng,et al.Brain Imaging&Behavior 2015,2018;IEEE TBME 20

21、15;Neuroinformatics 2017)Multi-Atlas Classification(M.Liu,D.Zhang,et al.Human Brain Mapping,2015;IEEE TMI 2016;TBME 2016)Illustration of multi-center adaptation with low-rankrepresentation learning for disease diagnosis(M.Wang,D.Zhang and etc.,MICCAI 2018)Multi-Center Disease ClassificationLow-Rank

22、Representation Method(M.Wang,D.Zhang and etc.,MICCAI 2018)Imaging GeneticsAssociationTree Structure Among SNPs(a)group by gene(b)group by linkage disequilibrium(LD)clusterTree-Guided Sparse LearningTGSL.SNP2SNP4SNPiSNP1.SVRTree Construction by Gene or L D Hierarchical ClusteringSNPsSelectionSNP1SNP2

23、SNP3SNP4.SNPiSNPjSNPkSNPkMRISimultaneous feature selection and regrTGSL(X.Hao,et al.MICCAI,2014;IEEE TCBB 2018)Multi-modality Phenotype Associations(X.Hao,et al.PSB,2016;Neuroinformatics 2016)Multi-SNP-Multi-QT Associations(X.Hao,et al.Scientific Reports 2017)wrw3=0v3x1u1x2u2=0 x3u3x4u4=0 xpupw1=0w2

24、.v1=0y1v2=0y2y3vqyq.XYz1z2x3.zrZLongitudinal Phenotype Associations(X.Hao,et al.Bioinformatics 2017)Brain Decoding:HyperalignmentBrain Decoding:HyperalignmentLocal Discriminant Hyperalignment(Yousefnezhad&Zhang,AAAI 2017)Original SpaceCommon Space()Subject 1Subject SVoxel x2VoxelVoxelVoxel x2Deep Hy

25、peralignment(Yousefnezhad&Zhang,NIPS 2017)OutlineSummary123Backgrounds on Alzheimers DiseaseBrain-imaging based AnalysisBrain-network based Analysis4Brain Connectomics Studies the interaction of brain functional regions at systems level(Petra E.Vrtes,et al.,PNAS,2012)Mapping the human brain is one o

26、f the great scientific challenges of the 21st centuryConnectivity Analysis(Honey et al.,PNAS,2007)Network-based ClassificationNeuroimaging dataNetworks constructionFeature extraction and selectionTraining classifierNetwork-based ClassificationMotivation Brain connectivity networks have been used for

27、classification of AD/MCI from normal controls(NC)In conventional methods,local measures of connectivity networks are first extracted from each ROI as network features,and then concatenated into a long vector Some useful structural information of network,especiallyglobal topological information,may b

28、e lostQuestion:How can we better preserve the network topological information for more effective brain network based classification?Topological Graph Kernel Topology-based graph kernel The kernel is defined on graphs,which can be used to computethe similarity of a pair of graphsfMRI imageRegional me

29、an time seriesConnectivity networkGray matter maskT1T2TMRFE-GKRFE-GKRFE-GKKernel combinationKernel matrixKernelmatrixKernelmatrix0.40.20-0.2-0.4-0.60.80.6t-testt-testt-testSVM Classifier0.90.80.70.60.50.40.30.20.1T10.90.80.70.60.50.40.30.20.10.90.80.70.60.50.40.30.20.1T2TMThresholded connectivity ne

30、tworksFF eeaattuurr eeextractionRFE-GKRFE-GKRFE-GKt-testt-testt-testFFeeaattuurree selectionFlowchart(B.Jie,D.Zhang,et al.,Human Brain Mapping,2014)Classification resultsMethodsT1T2T3T4T5combinedACC(%)VEC-RFE-LK-83.8VEC-RFE-RBF-73.0t-test75.778.464.964.964.981.1RFE-RBF78.473.067.673.078.486.5RFE-LK8

31、3.870.364.978.464.986.5RFE-GK86.583.875.775.764.991.9SEN(%)VEC-RFE-LK-83.3VEC-RFE-RBF-66.7t-test75.075.050.050.050.083.3RFE-RBF58.366.725.033.350.083.3RFE-LK91.758.341.766.750.091.7RFE-GK91.775.058.366.750.0100.0SPE(%)VEC-RFE-LK-84.0VEC-RFE-RBF-76.0t-test76.080.072.072.072.080.0RFE-RBF88.076.088.092

32、.092.088.0RFE-LK80.076.076.084.072.084.0RFE-GK84.088.084.080.072.088.0AUCVEC-RFE-LK-0.85VEC-RFE-RBF-0.79t-test0.840.860.740.710.680.86RFE-RBF0.680.770.750.650.760.83RFE-LK0.870.820.700.790.720.89RFE-GK0.850.860.770.780.600.94ROC curve0.10.21000.10.20.30.410.90.80.70.60.50.30.40.50.60.70.80.9False po

33、sitive rate(1-Specificity)True positive rate(sensitivity)ROC curveRandom VEC+RFE-LK VEC+RFE-RBFt-test Combined RFE-RBF Combined RFE-LK Combined RFE-GK CombinedBrain Sub-networksT3(a)NC(b)MCIThresholded average connectivity sub-network basedon top selected ROIsT2T1(B.Jie,et al.IEEE TBME,2014)(B.Jie,e

34、t al.IEEE TIP,2018)Sub-network Kernel Based MethodHyper-network based MethodMotivation Conventional connectivity network is usually constructed based on the pairwise correlation among brain regions Cannot reflect the useful higher-order relationship among brain regionsQuestion:how to character the h

35、igher-order relationship among brain regions?Solution:Hyper-graphV1V3V2V4V5V6V7=(,)=,=,=,=,=,=,VV7V6e15e3V3e2e4V2V4 V1 Hyper-graph vs.graph.Left:a conventional graph in which two nodes are connect-edtogether by an edge.Middle:a hyper-graph in which each hyper-edge can connect morethan two nodes.Righ

36、t:the incidence matrix for the hyper-graph in the middleFlowchart(B.Jie,D.Shen,D.Zhang,MICCAI14;Medical Image Analysis 2016)Experimental ResultsClassification performances of different methodsMethodAccuracySensitivitySpecificityAUCCN_CC HN_HCC1HN_HCC2 HN_HCC3Proposed62.275.781.189.294.641.741.775.08

37、3.391.772.092.084.092.096.00.540.750.800.930.96Results0.10.20.80.91000.10.20.30.40.510.90.80.70.60.30.40.50.60.7False positive rateTrue positive rateROC curveCN_CC HN_HCC1 HN_HCC2 HN_HCC3ProposedHyperedges(c)L.anterior cingulate gyrus(L.ACG)NCMCI(d)L.middle cingulate gyrus(L.MCG)NCMCI(a)L.middle fro

38、ntal gyrus(L.MFG)(b)L.rectus gyrus(L.REC)Connectome Biomarkers(C.Zu,et al.Brain Imaging and Behavior 2018)Top 10 Subnetworks(C.Zu,et al.Brain Imaging and Behavior 2018)Ordinal Patterns MiningExisting network descriptors Node degrees Clustering coefficients Sub-networks Limitations of previous work D

39、esigned on un-weighted brain connectivity networks Focus on individual brain regions other than local structures of brain networksAn overview of ordinal pattern based learning for brain disease diagnosis(X.Liu,et al.,MICCAI 2016;D.Zhang,et al.,IEEE TMI 2018)FlowchartIllustration of the proposed ordi

40、nal patternsIllustrationExperimental ResultsMethodAD vs.NCMCI vs.NCADHD vs.NCACCSENAUCACCSENAUCACCSENAUCCC72.6273.5370.9471.1472.7368.6971.2972.0370.51CCMT80.9582.3576.3574.5075.7674.7974.5375.4377.64DS76.1976.4775.5977.1878.7974.8981.0181.3680.82DSMT85.7185.2987.5979.1980.8176.9983.7984.7484.63Proposed94.0596.7796.3588.5987.2784.5787.5088.8987.37Comparison of different methods in three classification tasks(a)Top 2 ordinal patters from ADHD(b)Top 2 ordinal patters from NCOrdinal patterns identified in AD vs.NC classificationIdentified Ordinal PatternsThanks for your attention!